AI Agents Learn Their Way Up in Adversarial Game Agent Competition
ORIGINAL / Can AI Agents Learn Their Way to the Top? Evaluating Heuristic Learning in a Long-Running Game Agent Competition
This study formalizes Adversarial Heuristic Learning (AHL), where AI agents iteratively revise game strategies and supporting software while model weights stay fixed, evaluated on AAArena, a benchmark of 12 authentic adversarial games and 1,920 archived human programs. Its novelty lies in treating agents as learning engines rather than relying solely on reinforcement learning; it is useful because it shows opponent selection and dense feedback can improve policies, yet half o
01 ABSTRACT
The paper treats AI agents as learning engines that improve adversarial game performance by analyzing replays and revising executable policies while keeping model weights fixed, formalized as Adversarial Heuristic Learning (AHL). The authors built AAArena, a benchmark of 12 authentic adversarial games and 1,920 archived human programs, with an evaluation protocol modeled on real competitions. Experiments show Opus5.5 with Claude Code earns 6 gold medals, but no evaluated configuration tops the remaining 6 human ladders, and performance is weaker in games with more complex rules.
02 KEY FINDINGS
- Introduces Adversarial Heuristic Learning (AHL), where agents revise game policies and supporting software while model weights stay fixed.
- Builds AAArena benchmark with 12 adversarial games and 1,920 archived human programs under a competition-style evaluation protocol.
- Opus5.5+Claude Code earns 6 gold medals but fails to top 6 human ladders; opponent selection and dense feedback support policy improvement.
AI GENERATED SUMMARY / DISCOVERED BY ARXIV CS.CL